RoseTTAFold

What it is

RoseTTAFold is a method for predicting protein structures with a three-track neural network. The paper describing it appeared in Science on 15 July 2021, from a team in the Baker lab led by Minkyung Baek.^[1][2] It works from sequence, and covers both the shape of a single protein and the way proteins assemble into complexes.

The three tracks

A protein is a chain of amino acids, and predicting its structure means reasoning about several kinds of information at once. RoseTTAFold arranges this into three tracks: the 1D level of the amino acid sequence, the 2D level of a distance map recording how far apart pairs of residues sit, and the 3D level of coordinates in space. Information at these three levels is successively transformed and integrated, so that what one track represents constrains the others, and results are passed back and forth rather than treated as separate problems.^[1] Because the tracks are connected, the network can use constraints that reach across a whole chain while it builds a structure, instead of deciding sequence features and geometry independently.

Accuracy at CASP14

At CASP14, the community's blind assessment in which methods predict structures that have just been determined experimentally, RoseTTAFold's predictions approached those of DeepMind's AlphaFold2.^[1] Being close to AlphaFold2 on the same set of targets meant that groups had a second method whose accuracy was in the same range.

What it is used for

The method enables rapid solution of challenging modelling problems in X-ray crystallography and cryo-electron microscopy.^[1] It also generates models of protein–protein complexes from sequence alone, short-circuiting the usual route of modelling each subunit first and then docking the pieces together.^[1] According to the Institute for Protein Design, a structure can be computed in as little as ten minutes on a single gaming computer.^[2] The team used it to compute hundreds of new structures, including poorly understood human proteins and targets in lipid metabolism, inflammation and cancer cell growth.^[2]

Access, licence and requirements

A public web server was part of the release, and more than 4,500 proteins were submitted to it in its first month.^[2] The code is released under the MIT licence, while the trained weights are for non-commercial use under the Rosetta-DL Software licence.^[3] Running the method locally requires large reference databases: UniRef30 at about 46 GB, BFD at about 272 GB, and a PDB100 template set of over 100 GB.^[3]

Later versions

In November 2021 a two-track variant, RF2t, was added for yeast protein–protein interaction screening.^[3]

Where it sits

RoseTTAFold combines several things that are usually separate: accuracy that approached AlphaFold2 at CASP14, models of protein–protein complexes built from sequence alone, code released under an open licence, and a public web server that took thousands of submissions in its first month.^[1][2][3] Together these made structure prediction something a lab could run itself and build on.